Alibaba-NLP/UEmbed-2B

UEmbed: Unified Sparse and Dense Multimodal Embeddings

UEmbed is a decoder-only multimodal embedding model that produces both dense embeddings and SPLADE-style sparse lexical embeddings from a single causal forward pass. It supports text, image, video, and mixed-modal inputs for retrieval, multimodal search, and visual-document retrieval.

Model Family

Model	Backbone	Parameters	Outputs	Modalities
UEmbed-2B	Qwen3.5	2B	Dense + Sparse	Text, image, video
UEmbed-4B	Qwen3.5	4B	Dense + Sparse	Text, image, video
UEmbed-9B	Qwen3.5	9B	Dense + Sparse	Text, image, video

Highlights

• Unified dense and sparse retrieval: one checkpoint returns normalized dense vectors and sparse lexical vectors.
• Multimodal inputs: text, images, videos, and mixed inputs are represented in the same retrieval space.
• Sparse interpretability: sparse activations correspond to vocabulary terms and can be used with inverted indexes.
• Causal-model serving compatibility: the sparse design keeps the decoder-only backbone, no conversion to a bidirectional encoder.

Architecture

Component	Design
Backbone	Decoder-only Qwen3.5 multimodal model
Dense pooling	Hidden state of the EOS token before sparse special tokens
Sparse tokens	N=16 appended special tokens
Sparse heads	One subset-specific linear head per special token
Sparse vocabulary	Compressed from 248,320 tokenizer entries to 184,016 canonical entries
Sparse activation	log(1 + ReLU(logits))
Training objective	Dense InfoNCE + sparse InfoNCE + query/document FLOPS regularization

Usage

Requires a recent transformers build with Qwen3.5/Qwen3-VL support:

pip install "transformers>=5.4.0" torch qwen-vl-utils tokenizers huggingface-hub pillow numpy

Download the complete model repository, since sparse inference requires both sparse_info.json and sparse_weights.pt in the local model directory:

huggingface-cli download Alibaba-NLP/UEmbed-2B --local-dir ./models/UEmbed-2B

Inference code is provided in the GitHub repository. Set pooling="last.normal" for dense embeddings or pooling="splade.last" for sparse embeddings.

import torch
from src.models.qwen35_embedding import Qwen35Embedder

model = Qwen35Embedder(
    model_name_or_path="./models/UEmbed-2B",
    torch_dtype=torch.bfloat16,
    # flash_attention_2 for better acceleration and memory saving
    attn_implementation="flash_attention_2",
)

inputs = [{
    "text": "A woman playing with her dog on a beach at sunset.",
    "instruction": "Retrieve images or text relevant to the user's query.",
}, {
    "text": "A woman shares a joyful moment with her golden retriever on a sun-drenched beach at sunset, as the dog offers its paw in a heartwarming display of companionship and trust."
}, {
    "image": "https://qianwen-res.oss-cn-beijing.aliyuncs.com/Qwen-VL/assets/demo.jpeg"
}, {
    "text": "A woman shares a joyful moment with her golden retriever on a sun-drenched beach at sunset, as the dog offers its paw in a heartwarming display of companionship and trust.",
    "image": "https://qianwen-res.oss-cn-beijing.aliyuncs.com/Qwen-VL/assets/demo.jpeg"
}]

embeddings = model.process(inputs)
print(embeddings @ embeddings.T)

Input Format

Qwen35Embedder.process accepts a list of dictionaries with the following fields:

Field	Type	Description
text	str or list[str]	Text content.
image	path, URL, PIL.Image, or list	One or more images.
video	path, URL, frame list, or list	One or more videos.
instruction	str	Optional task-specific instruction.
fps	float	Optional frame sampling rate for video files.
max_frames	int	Optional maximum number of sampled video frames.

Training Data

UEmbed is trained on 3.94M public samples:

• E5 training data for broad text retrieval coverage.
• M3 training data, using the MLDR subset.
• MMEB training sets for multimodal query-document pairs.

For multimodal data, hard negatives are mined with Qwen3-VL-Embedding-8B as the teacher retriever.

Citation

If you use UEmbed, please cite the paper:

@misc{uembed2026,
      title={UEmbed: Unified Sparse and Dense Multimodal Embeddings}, 
      author={Tingyu Song and Mingxin Li and Yanzhao Zhang and Dingkun Long and Pengjun Xie and Zhijie Nie and Yilun Zhao and Shu Wu},
      year={2026},
      eprint={2608.02583},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2608.02583}, 
}

Acknowledgements

Thanks to the Qwen3-VL-Embedding repo for the evaluation framework.